Continuous Symplectic Phase Spaces & Cache-Local AI Runtimes
Project AFT_Ouroboros investigates non-dissipative recurrent sequence processing, Clifford spinor representations, cellular cohomology memory crystallization, and deterministic execution strictly confined within L2 cache on edge silicon.
Core Research Vectors
Eliminating quadratic attention bottlenecks and memory fragmentation through rigorous mathematical physics and deterministic low-level engineering.
Symplectic Phase Spaces
Unitary recurrent evolution on hyperspheres S^{D-1} and real projective spaces RP^{D-1}. Non-linear state transitions governed by exact Grassmann-Plücker syzygies and dyadic Zak butterfly reductions.
Cellular Knot Crystallization
Hodge-theoretic orthogonal decoupling on simplicial complexes. Dynamic states crystallize into harmonic 1-cycles with zero KV-cache growth, yielding over 342× memory reduction across sequence horizons.
Cache-Confined Execution
Strict C99 and ARM64 NEON vectorized kernels designed for deterministic execution entirely within L1D/L2 cache (Snapdragon 870 / Kryo 585 Prime) without dynamic memory allocations or DRAM traffic.
Physical Telemetry & Evidentiary Benchmarks
Empirical measurements captured via high-resolution monotonic clocks and CPU-bound affinity on consumer mobile hardware.
| Evaluated Property | Target Platform | Measured Value | Epistemic Status |
|---|---|---|---|
| Topological Knot Crystallization | Kryo 585 Prime @ 3.19 GHz | 7.029 µs / knot | MEASURED |
| Recurrent Training Throughput (D=64) | Kryo 585 Prime (L2-Confined) | 97,747 tokens/sec | MEASURED |
| Recurrent Training Throughput (D=128) | Kryo 585 Prime (L2-Confined) | 48,285 tokens/sec | MEASURED |
| Context Length Scaling (L ∈ [64, 16384]) | C99 ARM64 NEON Pipeline | α = 1.00226 (R² = 0.99999) | MEASURED |
| Zero-KV Memory Compression | Multi-scale Knot Storage vs KV | ≥ 342× reduction | DERIVED |
| Zero-Copy Checkpoint Mapping Latency | Linux mmap / 64-byte aligned | 21.41 µs (0 bytes heap) | MEASURED |
Open Repositories & Artifacts
Reproducible codebases, hardware definitions, and empirical data published under open licenses.
aft-ouroboros-silicon-benchmarks
Reproducibility suite, anti-DCE benchmarking harnesses, assembly verification, and canonical evidence matrix on consumer silicon.
View repository →
AFT-Ouroboros RISC-V SoC
Open hardware linear sequence coprocessor for resource-constrained RISC-V SoCs defined in SystemVerilog.
View repository →
ouroboros-ns-snn
Navier-Stokes analytical framework for spiking neural networks with logarithmic number systems (LNS) and Mitchell multiplierless arithmetic.
View repository →
Scientific Inquiries
Direct all technical correspondence, reproducibility inquiries, and collaboration requests to the lead investigator.
Esteban Natanael González
Lead Researcher & Systems Architect • Project AFT_Ouroboros